AI-Enabled Cyber-Physical Systems: Convergence of Smart Sensors, Predictive Analytics, and Autonomous Control — A Structured Review of Architectures, Benchmark Evidence, and Deployment Challenges
Abstract
ABSTRACT
Cyber-physical systems (CPS) tightly couple computation with physical processes, and the integration of artificial intelligence is transforming them from rule-driven automation into adaptive, self-optimising infrastructure. This review examines the convergence of three technology layers — smart sensors with edge intelligence, predictive analytics for prognostics and forecasting, and autonomous control based on model predictive control (MPC) and reinforcement learning (RL) — and evaluates how well the layers function as one closed loop. A structured, PRISMA-informed review of 32 peer-reviewed and benchmark-defining studies published between 2007 and 2021 was conducted, and a layered reference architecture was derived from it. Published remaining-useful-life (RUL) results on the NASA C-MAPSS benchmark show that deep convolutional models reduced root-mean-square error (RMSE) from 37.56 cycles for a multilayer perceptron to 12.61 cycles on subset FD001 (a 66% reduction), and by 9–22% relative to recurrent models across all four subsets. A qualitative appraisal of five control paradigms indicates that learning-based MPC and shielded safe RL offer the strongest balance of constraint satisfaction, adaptivity, and safety assurance. Persistent obstacles include edge resource limits, distribution shift, adversarial vulnerability, and the absence of certification-grade assurance. A research agenda built on digital twins, federated learning, and verifiable learning-enabled control is proposed.
KEYWORDS: Cyber-physical systems, Artificial intelligence, Smart sensors, Edge computing, Predictive maintenance, Remaining useful life, Model predictive control, Reinforcement learning, Digital twin, Safe autonomy
Cyber-physical systems (CPS) tightly couple computation with physical processes, and the integration of artificial intelligence is transforming them from rule-driven automation into adaptive, self-optimising infrastructure. This review examines the convergence of three technology layers — smart sensors with edge intelligence, predictive analytics for prognostics and forecasting, and autonomous control based on model predictive control (MPC) and reinforcement learning (RL) — and evaluates how well the layers function as one closed loop. A structured, PRISMA-informed review of 32 peer-reviewed and benchmark-defining studies published between 2007 and 2021 was conducted, and a layered reference architecture was derived from it. Published remaining-useful-life (RUL) results on the NASA C-MAPSS benchmark show that deep convolutional models reduced root-mean-square error (RMSE) from 37.56 cycles for a multilayer perceptron to 12.61 cycles on subset FD001 (a 66% reduction), and by 9–22% relative to recurrent models across all four subsets. A qualitative appraisal of five control paradigms indicates that learning-based MPC and shielded safe RL offer the strongest balance of constraint satisfaction, adaptivity, and safety assurance. Persistent obstacles include edge resource limits, distribution shift, adversarial vulnerability, and the absence of certification-grade assurance. A research agenda built on digital twins, federated learning, and verifiable learning-enabled control is proposed.
KEYWORDS: Cyber-physical systems, Artificial intelligence, Smart sensors, Edge computing, Predictive maintenance, Remaining useful life, Model predictive control, Reinforcement learning, Digital twin, Safe autonomy